Active Learning for PolSAR image classification

One of the biggest problems, when supervised learning techniques are used, for training classifier, is the necessity of a big amount of labelled samples, including the problems and costs of carry out the labelling of the prototypes needed. SAR images are difficult to label due to the speckle noise, which increases the normal effort needed for labelling a normal image. In order to reduce the number of samples needed for the training process and do not lose accuracy in the classification processes, active learning techniques appears. The main goal of active learning is to reduce the size of the training labelled sets used. For this purpose active learning utilizes techniques based on the amount of information, the active learning algorithms attempts to find the most informative samples from the unlabelled set of available samples. The use of active learning for this purpose is used in other different works as. In this project it has been used, to carry out the active learning, the strategy of Query by Committee. Furthermore it has been utilized techniques based on the entropy to measure the amount of information that the samples can provide. To develop the active learning process two images from two different cities are used. The active learning has been employed to select training sets and then training a classifier using the prototypes selected by means of the active learning process. The classifiers used to test the performance of active learning are Neural Networks and Support Vector Machines, furthermore to test the performance of active learning also training sets generated using passive learning have been used for training the classifiers and compare the results achieved between the passive and active learning. The results obtained show that the active learning is a good manner to reduce the number of labels needed to classify SAR images or at least it is capable to obtain better results using the same amount of samples as the passive learning, for the classification of SAR images.Gonzálvez García, P. (2013). Active Learning for PolSAR image classification. http://hdl.handle.net/10251/32154.Archivo delegad